pi-compact-context

Let the AI compact your context — just ask in plain English.

pi's built-in /compact command can only be typed by you. If you ask the model to "compact the conversation", it can't: compaction isn't a tool it can call. This package fixes that. It gives the model a compact_context tool, so natural-language requests like these just work:

"compact context"

"compact context after pushing the docs"

"summarize what we've done and compact the session"

"we're done with the debugging phase — compact, then start fresh on the API"

The model calls the tool itself, the current turn finishes, and pi compacts the conversation exactly as if you'd typed /compact.

Install

pi install npm:pi-compact-context

or straight from git:

pi install git:github.com/aerialcombat/[email protected]

Then restart pi (or run /reload). That's it — no configuration.

What it adds

| Interface | Name | Description | |-----------|------|-------------| | AI tool | compact_context | Callable by the model. Optional instructions parameter shapes the compaction summary. | | Slash command | /compact-context [instructions] | Same behavior, typed by you. Handy alias if you don't like pi's built-in /compact spelling. |

Both queue the same compaction; results appear via a toast notification (Compacted: N → M est. tokens).

How it works (and why it returns immediately)

Compaction in pi must not run mid-turn: AgentSession.compact() aborts the running agent turn before it starts. If a tool called it synchronously from inside that very turn, the run and the compaction would kill each other — the tool deadlocks, and you get a stuck "No result provided".

So this extension is deliberately fire-and-forget:

1. You say "compact context".
2. The model calls compact_context.
3. The tool queues the compaction and returns instantly ("Compaction queued — it will run as soon as this turn ends").
4. The turn finishes normally.
5. pi compacts the now-idle session and shows the token counts.

You keep one full turn of headroom for the model to wrap up whatever it was doing — like pushing the docs — before the history gets summarized.

Requirements

• pi coding agent
• Any model — compaction uses pi's normal summarization path

FAQ

Does this replace /compact? No. /compact still works exactly as before. This adds an AI-triggerable path on top.

What about auto-compaction? Unaffected. pi still compacts automatically near the context-window threshold; this package only adds manual, conversational control.

Can the model compact itself mid-task and keep going? Not in the same turn — compaction needs an idle agent. The model should finish its current work, then call the tool. This is a safety property of pi's compaction, not a limitation of this package.

Where does the summary come from? pi's built-in compaction: an LLM summarizes the older history, recent messages are kept verbatim. Pass instructions (or type /compact-context focus on the API migration) to steer what the summary emphasizes.

License

MIT